PADDLES: Phase-Amplitude Spectrum Disentangled Early Stopping for Learning with Noisy Labels

Huaxi Huang, Hui Kang, Sheng Liu, Olivier Salvado, Thierry Rakotoarivelo, Dadong Wang, Tongliang Liu

Keywords: Learning with noisy labels, Frequency domain decomposition, Early Stopping Training

TL;DR: We propose a new early stopping training method for learning with noisy labels by choosing different stopping points for the Phase and Amplitude spectrum in the frequency domain.

Abstract: Deep Neural Networks (DNNs) have demonstrated superiority in learning various patterns. However, DNNs are sensitive to label noises and would easily overfit noisy labels during training. The early stopping strategy averts updating DNNs during the early training phase and is widely employed as an effective method when learning with noisy labels. Motivated by biological findings that the amplitude spectrum (AS) and phase spectrum (PS) in the frequency domain play different roles in the animal's vision system, we observe that PS, which captures more semantic information, is more resistant to label noise than AS. Performing the early stopping on AS and PS at the same time is therefore undesirable. In contrast, we propose early stops at different times for AS and PS. In order to achieve this, we disentangle the features of some layer(s) into AS and PS using Discrete Fourier Transform (DFT) during training. The AS and PS will be detached at different training stages from the gradient computational graph. The features are then restored via inverse DFT (iDFT) for the next layer. We term the proposed method Phase-AmplituDe DisentangLed Early Stopping (PADDLES). Simple yet effective, PADDLES outperforms other early stopping methods and obtains state-of-the-art performance on both synthetic and real-world label-noise datasets.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Unsupervised and Self-supervised learning

Community Implementations: [ 1 code implementation](https://www.catalyzex.com/paper/paddles-phase-amplitude-spectrum-disentangled/code)